We present a new method of variational data assimilation (VAR) using the quasi-inverse linear approach, which does not require either adjoint models or minimization algorithms. This method shows much faster convergence towards the minimum of cost function than the adjoint 4D-VAR method. For a single time level, our method is equivalent to the Newton algorithm without the need to compute the Hessian. Applications to Burgers' equation and a 3D storm model (ARPS) will be presented.

Key words: quasi-inverse method, variational data assimilation, inverse model, adjoint model, storm prediction